Re: [PATCH] Enable auto-vectorization at O2 with very-cheap cost model.

[Hongtao Liu <crazylht@gmail.com>]

On Mon, Sep 6, 2021 at 5:42 PM Jakub Jelinek via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Mon, Sep 06, 2021 at 11:18:47AM +0200, Richard Biener wrote:
> > On Mon, Sep 6, 2021 at 10:47 AM liuhongt via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Hi:
> > >   As discussed in [1], most of (currently unopposed) targets want
> > > auto-vectorization at O2, and IMHO now would be a good time to enable O2
> > > vectorization for GCC trunk, so it would leave enough time to expose
> > > related issues and fix them.
> > >
> > >   Bootstrapped and regtested on x86_64-linux-gnu{-m32,}
> > >   Ok for trunk?
> >
> > It changes the cost model used when the user specifices
> > -O2 -ftree-vectorize which used 'cheap' before but now sticks to
> > 'very-cheap'.  I guess adjusting the cost model in process_options
> > might be possible when any(?) of the vectorizer flags were set
> > explicitly?
>
> process_options would mean it affects only the command line and not
> __attribute__((optimize ("O2", "ftree-vectorize")))
> etc.
> So, shouldn't it be instead done in default_options_optimization, somewhere
It seems default_options_optimization is before read_comline_options
which means it can't handle cmdline option -O2 -ftree-vectorize.

  default_options_optimization (opts, opts_set,
                                decoded_options, decoded_options_count,
                                loc, lang_mask, &handlers, dc);

  read_cmdline_options (opts, opts_set,
                        decoded_options, decoded_options_count,
                        loc, lang_mask,
                        &handlers, dc);


> among the
>   if (openacc_mode)
>     SET_OPTION_IF_UNSET (opts, opts_set, flag_ipa_pta, true);
>
>   /* Track fields in field-sensitive alias analysis.  */
>   if (opt2)
>     SET_OPTION_IF_UNSET (opts, opts_set, param_max_fields_for_field_sensitive,
>                          100);
>
>   if (opts->x_optimize_size)
>     /* We want to crossjump as much as possible.  */
>     SET_OPTION_IF_UNSET (opts, opts_set, param_min_crossjump_insns, 1);
>
>   /* Restrict the amount of work combine does at -Og while retaining
>      most of its useful transforms.  */
>   if (opts->x_optimize_debug)
>     SET_OPTION_IF_UNSET (opts, opts_set, param_max_combine_insns, 2);
> in there?
> Like:
>   /* Use -fvect-cost-model=cheap instead of -fvect-cost-mode=very-cheap
>      by default with explicit -ftree-{loop,slp}-vectorize.  */
>   if (opts->x_optimize == 2
>       && (opts_set->x_ftree_loop_vectorize
>           || opts_set->x_ftree_slp_vectorize))
>     SET_OPTION_IF_UNSET (opts, opts_set, fvect_cost_model_,
>                          VECT_COST_MODEL_CHEAP);
> Though, unsure if that will work with -O2 -ftree-vectorize which is an
> option without flag with EnabledBy on the other two options.
>
> Also, is:
> +    { OPT_LEVELS_2_PLUS, OPT_ftree_loop_vectorize, NULL, 1 },
> +    { OPT_LEVELS_2_PLUS, OPT_ftree_slp_vectorize, NULL, 1 },
> what we really want, isn't that instead:
> +    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_ftree_loop_vectorize, NULL, 1 },
> +    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_ftree_slp_vectorize, NULL, 1 },
> ?  I mean, for -Os vectorization even in very-cheap model I'd think it
> usually enlarges code size, and for -Og it is seriously harmful for
> debugging experience, especially when DWARF <= 5 doesn't have anything that
> would help debugging vectorized loops.
>
>         Jakub
>


--
BR,
Hongtao